Airport Taxi Demand Forecasting
Hourly time-series forecast for airport taxi demand using lag features, seasonality analysis, and temporal validation. Compares a seasonal baseline, random forest, and XGBoost against a stated RMSE target.
Choose the ZIP for a complete local setup. The notebook-only download requires the datasets and dependencies from that bundle.
Evaluation plan.
The revised workflow removes target leakage from rolling features and selects the model before final testing. The earlier RMSE of 42.9 is retired because the previous evaluation was optimistic.
Temporal
time-ordered training and validation
≤ 48
project target, not a claimed result
1 hour
previous observed orders are available
10%
last chronological portion of the data
What the project tries to solve.
Forecast the number of taxi orders in the next hour so airport driver supply can be positioned more effectively during demand spikes.
Status: Revised study project. Includes data, dependencies, and a smoke-test mode.
Notebook: airport_taxi_demand_forecasting.ipynb
The ZIP includes the notebook, README, pinned Python dependencies, a command-line runner, and all required CSV datasets.
An hourly demand estimate could help plan driver availability. A comparison against the previous day makes it clear whether a more complex model adds useful predictive value.
Demand changes by hour and over the season. I chose this problem to practice using only information available at forecast time and evaluating models in chronological order.
How I approached it.
Built lagged and calendar-based features from hourly order history.
Inspect training-period trends and hour-of-day averages before fitting models.
Compare a previous-day baseline, random forest, and XGBoost using the same validation period.
Used temporal validation rather than ordinary shuffled cross-validation to preserve forecasting integrity.
What I would improve next.
Add weather, flight schedules, and holiday indicators where those inputs are available at forecast time.
Extend the walk-forward evaluation to multiple seasons and quantify drift.
Package the best model into a simple forecast endpoint or dashboard mockup.